Robot Behavior Selection Using Salient Landmarks and Object-based Attention

This paper proposes a vision-based behavior selection system using biologically-inspired visual attention selection mechanisms, i.e., Bottom-Up attention and Top-Down attention. We adopt purely Bottom-Up attention selection to identify conspicuous regions for obtaining the salient landmarks, while propose an object-based Top-Down attention method using low dimensional task-relevant feature for searching target. The autonomous behavior selection system utilizes the salient landmarks and topological map for localization and navigation based on position prediction of matched landmark pairs. The proposed system is evaluated using several tasks in indoor and office environments for mobile robot. The applicability and the usefulness of the developed method are validated by the results obtained in this manner.